Virtual contrast-enhanced maximum intensity projections from high-b-value diffusion-weighted breast MRI: a...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Maximum intensity projections (MIPs) are crucial for rapid lesion detection in breast MRI using contrast-enhanced (CE) and diffusion-weighted imaging (DWI).
• Evaluating AI-based virtual CE subtraction MIPs as a feasible reading approach is essential for improving diagnostic efficiency.

Purpose of the Study:

• To assess the feasibility of AI-generated virtual abbreviated contrast-enhanced (VAbCE) subtraction MIPs for breast MRI lesion detection.
• To compare the diagnostic performance, image quality, and artifact presence of VAbCE MIPs against unenhanced (UnE) and abbreviated CE (AbCE) MIPs.

Main Methods:

• A retrospective study of 540 multi-parametric breast MRI exams (2017-2020) with multi-b-value DWI.
• A 2D U-Net model was trained to generate VAbCE subtractions from UnE images.
• Two radiologists evaluated lesion suspicion, image quality, artifacts, and lesion conspicuity for UnE, VAbCE, and AbCE images.

Main Results:

• Cancer detection rates were 90.0% (UnE), 91.4% (VAbCE), and 94.3% (AbCE).
• VAbCE MIPs demonstrated higher inter-rater agreement (Cohen κ=0.53) than UnE (0.39), comparable to AbCE (0.58).
• No significant differences in conspicuity, quality, or reading time were found between VAbCE and AbCE; VAbCE had fewer artifacts.

Conclusions:

• AI-based VAbCE breast MRI enhances inter-rater agreement and offers slightly improved sensitivity over UnE images, approaching AbCE sensitivity.
• VAbCE MIPs derived from neural networks enable rapid visual assessment, suggesting potential for screening applications.
• Further research is needed to fully explore the diagnostic capabilities of VAbCE breast MRI.